---
license: cc-by-4.0
language: en
tags:
- retrieval
- skill-extraction
- esco
- graded-relevance
configs:
  - config_name: queries
    data_files:
    - split: validation
      path: queries/validation.parquet
    - split: test
      path: queries/test.parquet
  - config_name: corpus
    data_files:
    - split: corpus
      path: corpus/corpus.parquet
  - config_name: qrels
    data_files:
    - split: validation
      path: qrels/validation.parquet
    - split: test
      path: qrels/test.parquet
---

# skill-normalisation-esco-graded

Graded-relevance annotations for surface skill terms (ESCO alt-labels) from
[`ESCO v1.1.0 skill-normalisation pairs`](https://esco.ec.europa.eu)
against the ESCO v1.1.0 skill taxonomy. Layout follows the
[BEIR](https://github.com/beir-cellar/beir) convention so it is drop-in for
MTEB-style retrieval evaluators.

This dataset was created for the RecSys-HR 2026 WorkRB challenge.

## Configs

| config | split | rows | columns |
|---|---|---:|---|
| `queries` | `validation` | 50 | `_id` (query id), `text` (ESCO alt-label / surface term to normalise) |
| `queries` | `test` | 450 | `_id` (query id), `text` (ESCO alt-label / surface term to normalise) |
| `corpus`  | `corpus` | 13,891 | `_id` (ESCO skill URI), `title` (English preferred label), `text` (English description), `esco_version` |
| `qrels`   | `validation` | 694,550 | `query-id`, `corpus-id`, `score` (0-4) |
| `qrels`   | `test` | 6,250,950 | `query-id`, `corpus-id`, `score` (0-4) |

Higher is more relevant. **The corpus is the COMPLETE ESCO v1.1.0 skill
list (13,891 skills)** The qrels are the full query x corpus grid:
every query is scored against all 13,891 skills, with score 0 for any skill
that was not surfaced as relevant.

### Score scale, `validation` split

| score | volume | meaning |
|---:|---:|---|
| 0 | 665,668 | The skill is totally unrelated to the query. |
| 1 | 28,438 | The skill's domain is correct. It's a plausible skill in a broader context, but not mentioned in this query. |
| 2 | 312 | The skill could be recommended, but its granularity makes it not core to the query. |
| 3 | 58 | The skill is strongly relevant for this query, although it is more implied than explicitly demonstrated. |
| 4 | 74 | The skill is explicitly demonstrated or requested by the query, and is therefore a clearly correct recommendation. |

### Score scale, `test` split

| score | volume | meaning |
|---:|---:|---|
| 0 | 5,989,638 | The skill is totally unrelated to the query. |
| 1 | 257,118 | The skill's domain is correct. It's a plausible skill in a broader context, but not mentioned in this query. |
| 2 | 2,723 | The skill could be recommended, but its granularity makes it not core to the query. |
| 3 | 966 | The skill is strongly relevant for this query, although it is more implied than explicitly demonstrated. |
| 4 | 505 | The skill is explicitly demonstrated or requested by the query, and is therefore a clearly correct recommendation. |

## Usage

```python
from datasets import load_dataset

queries = load_dataset("TechWolf/Skill-normalisation-ESCO-graded", "queries", split="test")
corpus  = load_dataset("TechWolf/Skill-normalisation-ESCO-graded", "corpus",  split="corpus")
qrels   = load_dataset("TechWolf/Skill-normalisation-ESCO-graded", "qrels",   split="test")
```

Join on ids: `qrels["query-id"]` matches `queries["_id"]`, and
`qrels["corpus-id"]` matches `corpus["_id"]` (a full ESCO skill URI).


## Attribution

This dataset uses the ESCO classification of the European Commission
(ESCO v1.1.0, <https://esco.ec.europa.eu>), licensed under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The ESCO content has
been extracted into a tabular subset (skill URI, English preferred label,
English description); no semantic modifications were made. The European
Commission is not responsible for any use of the data.

Source queries come from
[`ESCO v1.1.0 skill-normalisation pairs`](https://esco.ec.europa.eu)
(also CC BY 4.0). The judge labels themselves are released under CC BY 4.0.

## Citation
To be announced